An intelligent optimization method for the design parameters of non-standard cables

Abstract

The present invention relates to the field of data processing. More specifically, the present invention relates to an intelligent optimization method for non-standard cable design parameters. The method includes: obtaining training samples; obtaining a pre-constructed cable design network, inputting a training sample into the pre-constructed cable design network, and obtaining a loss value; calculating the sample reliability according to the loss value; taking the product of the sample reliability and the loss value as the corrected loss value, using the corrected loss value to reversely update the parameters in the cable design network, and sequentially inputting the training samples into the cable design network to train it until the corrected loss value converges and the training is completed; inputting the data of the materials required for the current cable design into the trained cable design network to obtain the predicted design parameter data, and using the predicted design parameter data to optimize the given design parameter data. By accurately evaluating the sample situation, the network learning process is controlled, thereby improving the prediction accuracy of the network.

Description

Technical Field

[0001] The present invention relates to the field of data processing. More specifically, the present invention relates to an intelligent optimization method for non-standard cable design parameters. Background Art

[0002] When conducting cable design, not only the current-carrying capacity, insulation performance, and environmental adaptability need to be considered, but also the economic cost. Therefore, to complete high-quality cable design, not only high technical capabilities but also rich design experience are required. Since the design experience or technical capabilities of some designers cannot meet the design requirements, the cable design parameters designed by them need to be optimized and adjusted.

[0003] Most cable design parameter optimization methods estimate a cable design parameter through a neural network, and then use the estimated cable design parameter as a reference to optimize the cable design given by the designer. Therefore, the accuracy of the estimated cable design parameter will affect the optimization accuracy. To improve the accuracy of the estimated cable design parameter, it is necessary to improve the prediction accuracy of the neural network. An important factor affecting the prediction accuracy of the neural network is the accuracy of the training samples. For example, the accuracy of the cable design parameters in some training samples is poor, and it is difficult to guarantee the prediction accuracy of the neural network trained by them. Therefore, how to control the training process of the neural network according to the accuracy of the training samples has become the research focus of the present invention.

[0004] The patent application document with the publication number CN116675182A discloses a reformer and its design parameter optimization method. The method in this patent application document verifies and optimizes the design parameters through simulation experiments. The method in this patent application document does not involve the content of "controlling the training process of the neural network according to the accuracy of the training samples", so the method in this patent application document cannot solve this technical problem well. Summary of the Invention

[0005] To solve the problem of how to control the training process of the neural network according to the accuracy of the training samples, the present invention proposes an intelligent optimization method for non-standard cable design parameters, which includes the following steps:

[0006] Obtain the data of the materials required for cable design and the corresponding design parameter data, and use the design parameter data as the label of the corresponding material data to obtain training samples;

[0007] Obtain the pre-constructed cable design network, input a training sample into the pre-constructed cable design network, and obtain the loss value; Calculate the sample reliability , represents the similarity between the th training sample and the similar training samples, represents the The similarity between the label of a training sample and the labels of similar training samples, Represents the sequence of similarities between all training samples and similar training samples, Represents a sequence of similarities between the labels of all training samples and the labels of similar training samples, represents the correlation between two series. represents normalization processing; the product of the sample reliability and the loss value is used as the corrected loss value, the corrected loss value is used to reversely update the parameters in the cable design network, and the training samples are sequentially input into the cable design network for training until the corrected loss value converges and the training is completed;

[0008] The data required for the current cable design is input into the trained cable design network to obtain estimated design parameter data, and the estimated design parameter data is used to optimize the given design parameter data. The optimized design parameter data is positively correlated with the estimated design parameter data and the given design parameter data.

[0009] The present invention evaluates the accuracy of each training sample, and controls the degree of learning of the cable design network on the training sample according to the accuracy of each training sample, thereby preventing the cable design network from learning too much erroneous training sample information and improving the prediction accuracy of the cable design network; further, when evaluating the accuracy of the training samples, the characteristic that the corresponding design parameters should be similar when the material data are similar is taken into account, and the accuracy of the training samples is evaluated according to the consistency relationship between the similarity of the training samples and the similarity between the labels, thereby improving the accuracy of the evaluation.

[0010] Preferably, the obtaining of a pre-built cable design network comprises:

[0011] The network structure of the cable design network is based on the ANN network, and an attention vector is added between the input layer and the hidden layer, and the data in the attention vector represents the attention to each input data;

[0012] A loss function of the constrained attention vector is constructed, and all training samples are input into the cable design network in sequence. The first stage training of the cable design network is completed using the loss function of the constrained attention vector and the loss function of the ANN network. The cable design network when the first stage training is completed is recorded as the pre-constructed cable design network.

[0013] The present invention obtains the attention paid to each data in the cable design process by adding an attention vector into the network, thereby providing a basis for subsequent accurate analysis of the consistency relationship between the similarity of training samples and the similarity between labels.

[0014] Preferably, the loss function for constraining the attention vector is constructed, including:

[0015] ;

[0016] in, Represents the loss value obtained when the attention vector does not participate in the calculation, Represents the loss value obtained when the attention vector participates in the operation, represents the variance of the data in the attention vector, Represents the loss function for constraining the attention vector.

[0017] When constructing the loss function, the present invention controls the network's learning of useful data information by introducing the change in the loss value of whether the attention vector participates in the calculation. At the same time, the variance of the data in the attention vector is introduced to adjust the difference in attention of different data information in the network. Therefore, the loss function constructed in this way can extract the attention information more accurately.

[0018] Preferably, the method for obtaining the similarity between the training sample and similar training samples includes:

[0019] Based on the information data in the training samples, the training samples are clustered and the training samples in the same category are regarded as similar training samples;

[0020] The difference between each data point in the training sample and the corresponding data point of a similar training sample is calculated and recorded as the first difference. The attention vector is obtained in the cable design network when the first stage of training is completed. Each data in the attention vector is used as the weight of the first difference. The first difference values ​​obtained from all data points in the training sample are weighted and summed, and then the inverse is calculated to obtain the individual similarity between the training sample and the similar training sample. The average of the individual similarities between the training sample and all similar training samples is recorded as the similarity between the training sample and the similar training samples.

[0021] When calculating the similarity of training samples, the present invention uses the data in the attention vector as weights, so that data with high attention has a greater decisiveness in the similarity calculation, and data with low attention has a smaller decisiveness in the similarity calculation, providing a basis for the subsequent accurate description of the consistency relationship between the similarity of training samples and label similarity.

[0022] Preferably, inputting a training sample into a pre-built cable design network to obtain a loss value comprises:

[0023] A training sample is input into the cable design network, and the output value of the cable design network is obtained with the participation of the attention vector in the calculation. Based on the output value and the label, the loss value is calculated using the loss function of the ANN network.

[0024] Preferably, the optimization of the given design parameter data using the estimated design parameter data includes:

[0025] Obtain test samples, input the test samples into the cable design network, calculate the loss values of the test samples, record the normalized data of the mean of the loss values of all test samples as the prediction error degree, multiply the reciprocal of the prediction error degree by the estimated design parameter data to obtain a prediction correction value, and multiply the given design parameter data by the difference between 1 and the prediction error degree and then add the prediction correction value to obtain the corrected design parameter data.

[0026] The present invention controls the correction ratio through prediction accuracy, thereby moderately correcting the design parameter data and improving the accuracy of design parameter optimization.

[0027] Preferably, the method for obtaining the correlation between the two sequences includes:

[0028] Take the absolute value of the Pearson correlation coefficient of the two sequences as the correlation between the two sequences.

[0029] Preferably, the step of sequentially inputting the training samples into the cable design network for training includes:

[0030] Sequentially input each training sample into the cable design network, calculate the corrected loss value corresponding to each training sample input, and use the corrected loss value to perform backpropagation update on the parameters in the cable design network.

[0031] The present invention has the following beneficial effects:

[0032] The present invention evaluates the accuracy of each training sample and controls the learning degree of the cable design network for the training sample according to the accuracy of each training sample, thereby preventing the cable design network from learning too much incorrect training sample information and improving the prediction accuracy of the cable design network;

[0033] Furthermore, when evaluating the accuracy of the training sample, considering the feature that the corresponding design parameters should be similar when the data is similar, the accuracy of the training sample is evaluated according to the consistency relationship between the similarity of the training sample and the similarity of the label, thereby improving the accuracy of the evaluation. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0035] Figure 1It is a flowchart of the steps of an intelligent optimization method for non-standard cable design parameters according to an embodiment of the present invention. Detailed implementation manners

[0036] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present invention.

[0037] Next, the detailed implementation manners of the present invention will be described in conjunction with the accompanying drawings.

[0038] Please refer to Figure 1 , which shows a flowchart of the steps of an intelligent optimization method for non-standard cable design parameters provided by an embodiment of the present invention. The method includes the following steps:

[0039] S1: Obtain the data of the materials required for cable design and the corresponding design parameter data, and use the design parameter data as the label of the corresponding material data to obtain a training sample.

[0040] Specifically, collect the data of the materials required for each cable design and the design parameter data given by the designer according to the material data. The data of the materials required for cable design include but are not limited to the following aspects: the rated voltage and current that the cable needs to carry, the cable installation method (overhead installation, buried installation), and the cost budget. The design parameter data includes but is not limited to the following aspects: conductor material, conductor cross-sectional area, and insulating material.

[0041] Use the corresponding design parameter data as the label of the data of the materials required for cable design, and use the data of the materials required for cable design with labels as the training sample.

[0042] S2: Obtain the pre-constructed cable design network, input a training sample into the pre-constructed cable design network to obtain a loss value; calculate the sample reliability; use the product of the sample reliability and the loss value as the corrected loss value, and use the corrected loss value to update the parameters in the cable design network in the reverse direction. Input the training samples into the cable design network in turn to train it until the corrected loss value converges and the training is completed.

[0043] S20: Obtain the pre-constructed cable design network, input a training sample into the cable design network to obtain a loss value; calculate the sample reliability.

[0044] It should be noted that in order to estimate the cable design parameters, it is necessary to first construct a cable design network for estimating the cable design parameters.

[0045] Optionally, as an example, a pre-built cable design network is obtained, a training sample is input into the cable design network, and a loss value is obtained; and sample reliability is calculated, including:

[0046] Design ANN networks as pre-built cable networks.

[0047] A training sample is input into the cable design network to obtain the output data of the cable design network; based on the output data and the corresponding label, the loss value is calculated using the loss function of the cable design network.

[0048] The sample reliability is set to 1.

[0049] It is understandable that when training the cable design network traditionally, each training sample is considered accurate, without considering that the design parameters given by the designer may be wrong. Therefore, when training the cable design network, the trust in each training sample is the same, and the traditional default sample reliability is 1.

[0050] It should be noted that in order to improve the accuracy of cable design network training, the accuracy of the labels of the training samples needs to be evaluated, so as to adjust the trust of the cable design network for each training sample according to the accuracy of the training samples.

[0051] It should be further explained that, under normal circumstances, when the data required for cable design are similar, the corresponding design parameters should be similar. Therefore, the reliability of the labels of training samples can be analyzed based on this logic. At the same time, for cable design, some data are more important, and some data are less important. Therefore, when designing cables, the attention paid to each data is different. In addition, the attention paid to the data will affect the accuracy of the label reliability evaluation of the training samples. For example, some important data are similar, and the corresponding design parameters should be similar. For some unimportant data, it is only used as a reference to fine-tune the design parameters. Because these unimportant data are different, they will not cause large differences in the design parameter data. Therefore, when using the similarity and consistency logic of data and design parameter data to evaluate the label reliability of training samples, the importance of the data needs to be considered.

[0052] Preferably, as an example, a pre-built cable design network is obtained, a training sample is input into the pre-built cable design network, and a loss value is obtained; and the sample reliability is calculated, including:

[0053] First, get a pre-built cable design network including:

[0054] The network structure of the cable design network is based on the ANN network, and an attention vector is added between the input layer and the hidden layer. The data in the attention vector represents the attention to each input data; the dimension of the attention vector is the same as the dimension of the input data.

[0055] Construct the loss function of constrained attention vector:

[0056]

[0057] in, It represents the loss value obtained when the attention vector does not participate in the operation. This value reflects the loss value of the output data obtained by the cable design network when there is no attention vector. It represents the loss value obtained when the attention vector participates in the operation, which reflects the loss value of the output data obtained by the cable design network when considering the information attention. It reflects the change in the error rate of the output data obtained by the cable design network before and after considering the information attention. The larger the value, the greater the improvement in the prediction accuracy of the cable design network by considering the information attention. Represents the variance of the data in the attention vector. This value is used to increase the difference in the attention of the cable design network to different information, so that the cable design network pays more attention to important information and less attention to unimportant information. Represents the loss function for constraining the attention vector.

[0058] All training samples are input into the cable design network in sequence, and the first stage training of the cable design network is completed using the loss function of the constrained attention vector and the loss function of the ANN network. The cable design network when the first stage training is completed is recorded as the pre-built cable design network.

[0059] Then, a training sample is input into the pre-built cable design network to obtain the loss value, including:

[0060] A training sample is input into a pre-built cable design network. When the attention vector is involved in the calculation, the output data of the pre-built cable design network is obtained. Based on the output data and the corresponding labels, the loss value is calculated using the loss function of the ANN network.

[0061] Finally, the sample reliability is calculated, including:

[0062] Based on the information data in the training samples, the training samples are clustered and the training samples in the same category are regarded as similar training samples;

[0063] The difference between the data of each training sample and the corresponding data of a similar training sample is denoted as the first difference. At the end of the first-stage training, an attention vector is obtained from the cable design network. Each data in the attention vector is used as the weight of the first difference. After weighted summation of the first differences obtained from all the data of the training samples and then taking the reciprocal, the individual similarity between the training sample and the similar training sample is obtained. The mean of the individual similarities between the training sample and all similar training samples is denoted as the similarity between the training sample and the similar training sample. For example, calculate the difference between the k-th data of a training sample and the k-th data of a similar training sample, denoted as the target first difference. The k-th data in the attention vector is used as the weight of the target first difference. Similarly, each first difference and the corresponding weight are obtained. After weighted summation of all the first differences using the weights and then taking the reciprocal, the individual similarity between the training sample and the similar training sample is obtained. The mean of the individual similarities between the training sample and all similar training samples is used as the similarity between the training sample and the similar training sample.

[0064] It can be understood that when calculating the similarity between training samples, this method takes into account the importance of the data. Larger weights are assigned to important data so that important data plays a greater role in similarity calculation, and smaller weights are assigned to unimportant data so that unimportant data plays a smaller role in similarity calculation. Thus, the similarity of the training samples calculated by this method can maintain a consistent relationship with the label similarity. For example, important data is more decisive for cable design parameters, and unimportant data is less decisive for cable design parameters. Therefore, when important data is similar, the possibility of similarity in cable design parameters is relatively large. Thus, the possibility that the similarity calculated by the weighted method is the same as the label similarity is relatively large.

[0065] The sample reliability satisfies the formula:

[0066]

[0067] Wherein, represents the similarity between the label of the -th training sample and the label of the similar training sample, represents the sequence composed of the similarities between all training samples and the similar training sample, represents the sequence composed of the similarities between the labels of all training samples and the labels of the similar training sample, represents the correlation between the two sequences, reflects the consistency between the similarity of the data of all training samples and the similarity of the labels. At the same time, this value also characterizes the overall consistency. Since the incorrect label data is relatively few, the overall consistency index obtained from all training samples can reflect the index data under the accurate label situation. Therefore, It can be used as a reference to judge the accuracy of the labels of each training sample. Reflects the The consistency between the similarity of the data of the training samples and the similarity of the labels, This value reflects the The larger the value, the smaller the deviation of the consistency between the similarity of the data of the training samples and the label similarity. The greater the degree to which the consistency between the similarity of the data of the training samples and the label similarity deviates from the overall consistency, the The greater the possibility that the label of the training sample is wrong, the greater the probability that the The smaller the reliability of a training sample, represents the hyperbolic tangent function, which is used for normalization. Indicates The sample reliability of training samples is They represent the minimum function and maximum function respectively, which are used to obtain the minimum and maximum values ​​in the data.

[0068] It should be added that the method for obtaining the loss value obtained when the attention vector does not participate in the calculation includes:

[0069] The input data is not calculated with the attention vector, but directly flows into the hidden layer of the cable design network for the next step of calculation and analysis to obtain the output data of the cable design network; based on the output data and the corresponding labels, the loss value is calculated using the loss function of the ANN network.

[0070] In addition, the method for obtaining the loss value obtained when the attention vector participates in the calculation includes:

[0071] After the input data is multiplied by the attention vector, it flows into the hidden layer of the cable design network for the next step of calculation and analysis to obtain the output data of the cable design network; based on the output data and the corresponding labels, the loss value is calculated using the loss function of the ANN network.

[0072] In addition, all training samples are sequentially input into the cable design network, and the first stage training of the cable design network is completed by using the loss function of the constrained attention vector and the loss function of the ANN network. The cable design network after the first stage training is completed is recorded as the pre-constructed cable design network, including:

[0073] Input a training sample into the cable design network, and use the loss value calculated by the loss function that constrains the attention vector and the loss value obtained when the attention vector participates in the calculation to perform backpropagation updates on the parameters in the cable design network respectively, completing the training of the cable design network with this training sample. Similarly, complete the training of the cable design network with all training samples to obtain the cable design network after the first stage of training, denoted as the pre-constructed cable design network.

[0074] It should be noted that performing backpropagation updates on the network according to the loss value is a prior art, and will not be elaborated here.

[0075] S21: Use the product of the sample reliability and the loss value as the corrected loss value, and use the corrected loss value to perform backpropagation updates on the parameters in the cable design network. Input the training samples into the cable design network in sequence for training until the training is completed when the corrected loss value converges.

[0076] Preferably, as an example, inputting the training samples into the cable design network in sequence for training includes:

[0077] Input each training sample into the cable design network in sequence, calculate the corrected loss value corresponding to each training sample input, and use the corrected loss value to perform backpropagation updates on the parameters in the cable design network.

[0078] It should be noted that performing backpropagation updates on the parameters in the cable design network using the loss value is a prior art, and will not be elaborated here.

[0079] S3: Input the data of the materials required for the current cable design into the trained cable design network to obtain the estimated design parameter data, and use the estimated design parameter data to optimize the given design parameter data. The optimized design parameter data is positively correlated with both the estimated design parameter data and the given design parameter data.

[0080] S30: Input the data of the materials required for the current cable design into the trained cable design network to obtain the estimated design parameter data.

[0081] S31: Use the estimated design parameter data to optimize the given design parameter data.

[0082] Optionally, as an example, using the estimated design parameter data to optimize the given design parameter data includes:

[0083] Use the sum of the estimated design parameter data and the given design parameter data as the optimized design parameter data.

[0084] It should be noted that when optimizing the given design parameter data by taking the average value, the optimization ratio is not controlled, and it is very easy to have problems of over-optimization or under-optimization in such an optimization ratio.

[0085] Preferably, as an example, optimizing the given design parameter data by using the estimated design parameter data includes:

[0086] Re-obtain the data required for cable design and the design parameters given by the designer according to the data. Use the design parameters as labels for the corresponding data, and use the re-obtained data required for cable design with labels as test samples.

[0087] Input the test samples into the cable design network, calculate the loss values of the test samples, record the normalized data of the average value of the loss values of all test samples as the prediction error degree, multiply the reciprocal of the prediction error degree by the estimated design parameter data to obtain the prediction correction value, and multiply the given design parameter data by the difference between 1 and the prediction error degree and then add the prediction correction value to obtain the corrected design parameter data.

[0088] It can be understood that the optimization ratio is controlled by using the prediction accuracy. When the prediction accuracy is high, the optimization ratio is increased, so that the optimized data contains more accurate data and improves the accuracy of the optimized data; when the prediction accuracy is low, the optimization ratio is decreased, so as to prevent incorrect data from being introduced into the optimized design parameters and improve the accuracy of the optimized design parameters.

[0089] So far, this embodiment is completed.

[0090] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent optimization method for the design parameters of non-standard cables, characterized in that, include: Obtain the data and corresponding design parameter data required for cable design, and use the design parameter data as labels for the corresponding data to obtain training samples; Obtaining a pre-built cable design network, including: the network structure of the cable design network is based on the ANN network, and an attention vector is added between the input layer and the hidden layer, and the data in the attention vector represents the attention to each input data; constructing a loss function constraining the attention vector, the formula is: , represents the loss value obtained when the attention vector does not participate in the calculation, Represents the loss value obtained when the attention vector participates in the operation, represents the variance of the data in the attention vector, A loss function representing a constrained attention vector; inputting all training samples into the cable design network in sequence, using the loss function of the constrained attention vector and the loss function of the ANN network to complete the first stage of training of the cable design network, and recording the cable design network after the first stage of training is completed as the pre-constructed cable design network; Input a training sample into a pre-constructed cable design network to obtain a loss value; calculate the sample reliability , represents the similarity between the th training sample and the similar training samples, represents the similarity between the label of the th training sample and the labels of the similar training samples, represents the sequence composed of the similarities between all training samples and the similar training samples, represents the sequence composed of the similarities between the labels of all training samples and the labels of the similar training samples, represents the correlation between the two sequences, represents normalization; use the product of the sample reliability and the loss value as the corrected loss value, and use the corrected loss value to update the parameters in the cable design network in reverse. Input the training samples into the cable design network in turn to train it until the corrected loss value converges and the training is completed; The data required for the current cable design is input into the trained cable design network to obtain estimated design parameter data, and the estimated design parameter data is used to optimize the given design parameter data. The optimized design parameter data is positively correlated with the estimated design parameter data and the given design parameter data.

2. The intelligent optimization method for non-standard cable design parameters according to claim 1, wherein The method for obtaining the similarity between the training sample and the similar training sample includes: Based on the information data in the training samples, the training samples are clustered and the training samples in the same category are regarded as similar training samples; The difference between each data point in the training sample and the corresponding data point of a similar training sample is calculated and recorded as the first difference. The attention vector is obtained in the cable design network when the first stage of training is completed. Each data in the attention vector is used as the weight of the first difference. The first difference values ​​obtained from all data points in the training sample are weighted and summed, and then the inverse is calculated to obtain the individual similarity between the training sample and the similar training sample. The average of the individual similarities between the training sample and all similar training samples is recorded as the similarity between the training sample and the similar training samples.

3. An intelligent optimization method for non-standard cable design parameters according to claim 1, characterized in that The step of inputting a training sample into a pre-built cable design network to obtain a loss value includes: A training sample is input into the cable design network, and the output value of the cable design network is obtained with the participation of the attention vector in the calculation. Based on the output value and the label, the loss value is calculated using the loss function of the ANN network.

4. An intelligent optimization method for non-standard cable design parameters according to claim 3, characterized in that, The step of optimizing the given design parameter data by using the estimated design parameter data includes: Get a test sample, input the test sample into the cable design network, calculate the loss value of the test sample, record the normalized data of the mean loss value of all test samples as the prediction error degree, multiply the inverse of the prediction error degree by the estimated design parameter data to obtain the prediction correction value, multiply the given design parameter data by 1 and the difference between the prediction error degree and the prediction error degree, and then add the prediction correction value to obtain the corrected design parameter data.

5. The intelligent optimization method for non-standard cable design parameters according to claim 1, wherein The method for obtaining the correlation between the two sequences includes: The absolute value of the Pearson correlation coefficient of two series is taken as the correlation between the two series.

6. The intelligent optimization method for non-standard cable design parameters according to claim 1, characterized in that The step of sequentially inputting the training samples into the cable design network to train it includes: Each training sample is input into the cable design network in turn, the corrected loss value corresponding to each training sample input is calculated, and the corrected loss value is used to reversely update the parameters in the cable design network.

Images